Rotorcraft virtual sensors via deep regression

Highlights

• Deep fully-connected neural networks perform high-dimensional and non-linear mapping.

• Architecture design and hyper-parameters are critical components to configure.

• Evolutionary algorithms are effective and systematic methods for neural network optimization.

• Robust and representative objectives are required to ensure appropriate learning.

• An automated HPC asynchronous workflow was instrumental for model generation.

Abstract

Raw sensor data containing high-fidelity information is highly desirable for valuable post-processing. We developed a machine learning model that performs deep regression to infer rotorcraft component vibration spectra from a few flight conditional indicators (CI). The model consists of a deep neural network of fully connected layers (DNN) that performs high-dimensional and non-linear multivariate regression to reconstruct raw accelerometer data. The network architecture hyperparameters were optimized using an evolutionary genetic algorithm (GA) that was more effective than random and manual search methods. The best GA design was further tuned to achieve spectrum reconstruction accuracies above 95% on validation datasets. An automated model generator workflow was developed to train and evaluate thousands of DNN designs using parallel asynchronous execution on a Cray XC50, which were monitored and studied. Finally, as a verification step of the DNN inference model operation and performance, a detailed sensitivity analysis was performed using a modified Sobol sampling technique to understand response behavior and limitations. The sensitivity analysis method utilized Dask-distributed across multiple nodes on our HPC to evaluate millions of generated samples in parallel.